New supercomputer record

[menoinjun]

So there is a new king of the supercomputers in Japan.

http://news.com.com/2100-1001-887717.html?legacy=cnet&tag=lthd

It cost between 350 and 400 million dollars to build!! My question is, why didn't they just buy a huge cluster of macs? It would be a heck of a lot more cost effective!! Example:

The new computer has 5,104 processors, and produces 35,000 gigaflops of computing power. Just cluster 2500 PowerMac G4 1ghz duals together!! It would produce 37,500 gigaflops, and have 5000 processors!! Better yet, the cost is only $7,500,999 dollars!! (2500 G4s @ $2999 + one 23" Cinema display. Not including cables) It seems to me that some of the extra money could be put into development to come up with some superfast program to calculate the warming thing that the machine is doing. Considering how fast the DNA processing program runs that was a big deal a month ago that was built custom for Mac OS X(sorry but I forget the name) this computer would fly!! I could have saved Japan about $342 million dollars!! Some people just don't spend money wisely...

Comments?

-Pete

 

[mac15]

Wow thats fast
I want one

 

[ejm625]

Let's do it... everyone daisy-chain your macs together to blow out this computer!

 

[eirik]

clustered powermacs

Master Po,

Would not the Ethernet or the Firewire cables in daisy chain overload with but a dozen or two PowerMacs?

If so, might I suggest a number of Gigabit Ethernet switches to join them all together. Even so, the backplanes of these Ethernet switches would have to be huge.

Grasshopper

PS Have you ever considered wearing sunglasses, Master Po. Your freakishly opaque eyes give me nightmares!

 

[Mr. Anderson]

e

It would take more than 2500 Macs etherneted together to match the Nec machine. All the processors on that most likely share the same bus or atleast have a direct connection. Besides, to have such a massively parallel processor, you need to have software that can handle it. OSX wouldn't be able to handle 5000 processors - how much RAM do you think it has? Much more than 1.5 Gig

 

[mymemory]

For what?

In less than six months is gonna be discontinued.

I really need something like that now that I'm doing renderings. I'm doing a one minute, 30fps, 640 x 480 animation in my pismo and is taking about 5 to 6 hours to render.

 

[Mr. Anderson]

Re: For what?

Originally posted by mymemory
I really need something like that now that I'm doing renderings. I'm doing a one minute, 30fps, 640 x 480 animation in my pismo and is taking about 5 to 6 hours to render.

I did a 100 hour render, really couldn't use my machine for a week. And that wasn't much of an animation. I've said it before, I really won't be happy until I can do realtime radiosity, maybe in 10 to 15 years....

 

[eirik]

Re: Re: For what?

Originally posted by dukestreet


I did a 100 hour render, really couldn't use my machine for a week. And that wasn't much of an animation. I've said it before, I really won't be happy until I can do realtime radiosity, maybe in 10 to 15 years....

You know, I think this thread is on the threshold of something very interesting: numbers, or rather benchmarks.

How many GFLOPS would be required to attain Dukestreet's realtime radiosity quest? How many GFLOPS can the current high-end Macs do? I'd love to see a table of some kind illustrating our wants with the needed capabilities.

Eirik

 

[Choppaface]

why can't the build something like that and sell processor time? too expensive?

 

[eirik]

Originally posted by Choppaface
why can't the build something like that and sell processor time? too expensive?

Well, I'm sure we don't want to bother building a monstrocity such as the referenced Japanese Supercomputer. I say 'monstrocity' because I suspect its purpose has as much to do with national pride as with productive purposes.

Nonetheless, the idea of a rendering farm is interesting. However, we would need to know some performance specifications requirements so that we get a sense for just how big of a cluster we would need for just one job and then for multiple jobs. Then, we could make some estimates about overhead and other direct costs to get a crude sense of the economics. So, what would a rendering farm have to be capable of doing and what would it have to consist of? How cheap does a rendering farm have to be to be economical?

You see, this is an example of why I wrote that this thread might lead to something very interesting: performance metrics that would empower us to conceptualize all sorts of things, ranging from the ideal PowerMac to a rendering farm.

Eirik